allow_new_relation defaults still empty seeded sessions when forgotten

[chouswei]

Summary

Unknown @EDG relations abort the seed/add batch and roll back. session_open(..., allow_new_relation=False) is the default; forgetting the flag leaves rows=0 and is easy for LLM tool callers to miss.

Evidence

• src/memnet_mcp/server.py session_open docs and allow_new_relation default; CHANGELOG notes seed batches rolling back on first unknown relation.
• tests/test_mcp.py covers allow_new_relation and rollback behaviour.
• Skills/docs repeatedly remind callers to pass allow_new_relation=true for novel/SysML seeds (mcp-memnet skill, SysML notes).

Impact

"Empty session" failures during bootstrap; agents proceed with a broken graph; hard to distinguish from map/seed typos without reading stderr carefully.

Suggested direction

Clearer MCP error when seed rolls back to zero rows; optional "register relations from seed EDG lines" mode; or fail session_open loudly with a dedicated @ERR code when seed is empty after intended seed_lines.

[chouswei]

Still valid: allow_new_relation defaults false on session_open / add; unknown @EDG relations can still empty a seed batch. Not addressed by Tier A / MutateGate work.

[chouswei]

Refresh (2026-08): Still reproducible on master; unrelated to v0.3.6 serve hardening (3c537c7).

[chouswei]

Status (Hatch 0.19.1). session_open(..., allow_new_relation=False) is still the default in parts/memnet-mcp/software/memnet_mcp/server.py. Product mutate is GQL CREATE/MATCH…SET with kinds in the open SCHEMA map (unknown_tag if missing).

Leftover add seed batches can still roll back on unknown EDGE types. Empty-session bootstrap remains a real LLM footgun. Paths in the body (src/memnet_mcp) are stale — use parts/memnet-mcp/.